AI and Signatures

While thinking about my fascination with signatures, a thought appeared in my mind. Since AI consists of a large collection of images, I wondered if AI could produce "generic signatures" or not. As such, I decided to run a little experiment.

This is an iron sword from OldSchool RuneScape. This image depicts perfectly what I consider to be the "generic medieval sword" signature. My goal for the experiment was simple: seeing if AI can replicate something that is close to this exact signature.

To do this, the website Perchance was used. The only reason I chose this service was because it did not require me to create an account, unlike mainstream AI services such as DALL-E.

The prompt "Generic medieval fantasy sword" with the "No style" option gave me the following:

Although that's definitively a sword, it's not very generic. I do agree that the image gives heavy "generic concept art you find on Pinterest" kind of vibes though.

The prompt "Generic medieval fantasy sword in the style of OldSchool RuneScape" with the "No style" option gave me the following:

That looks like a dagger rather than a sword. It also seems that the AI cannot really grasp what the "OldSchool RuneScape" style is.

The prompt "Generic medieval fantasy sword" with the "Anime" style option gave me the following:

That result surprised me because my prompt did not include "anime girl". Although the sword is technically not generic, its color palette technically is.

The prompt "Generic medieval fantasy sword item image" with the "Anime" style option gave me the following:

Again, I am not sure why the AI insists that my image must feature an anime girl. The sword almost looks generic, the only thing wrong is the handle (which is also hilariously long for some reason).

The prompt "Generic medieval fantasy sword item image" with the "Flat illustration" style option gave me the following:

For starters that's not a sword. It kind of looks like a weird dagger like seen earlier. It is also far from being generic in design.

The prompt "Iron sword from OldSchool RuneScape" with the "No style" option gave me the following:

Yeah, I am not sure what happened there lol. That helmet sure looks very inconvenient to wear.

The prompt "Iron sword from OldSchool RuneScape" with the "No style" option and the negative prompt "Character" gave me the following:

What. I am speechless. The AI certainly dropped the ball hard on this one. The image does not even feature a sword anywhere. That thing seems like a mashup between a Minecraft map and an unfinished Unity game map.

So in conclusion, AI seems incapable of retrieving the generic sword signature on demand. I think that this is due to the fact the model is looking at images that only contain a sword that were scrapped on the internet. In those sorts of images, the artist is much more likely to want to create something unique.

I find interesting how all the attempts were it got close featured an anime girl. I theorize that the people who drew the anime girls in the dataset focused their efforts on the girl and the sword was an afterthought which meant they were more likely to subconsciously use the generic signature.

That was an interesting experiment. To me this seems to reveal that AI can only remix ideas which are already remixed and doesn't have knowledge of the "source signature". Speaking of which, I wonder where they originate from. I doubt that all the people drawing generic swords have all played RuneScape at some point in their life. Perhaps this specific signature is rooted in historical sword design. I might investigate this later.

Written by manpaint on 12 March 2025.